
3.6 The target trial 39
as the target experiment or the target trial. When conducting the target trialThe target trial—or its logical
equivalents—has long been cen-
tral to the causal inference frame-
work. Dorn (1953), Wold (1954),
Cochran (1972), Rubin (1974), Fe-
instein (1971), and Dawid (2000)
used the concept. Robins (1986)
generalized it for time-varying
treatments.
is not feasible, ethical, or timely, we resort to causal analyses of observational
data. Generally, one can view these observational analyses as an attempt to
emulate some target trial. If the emulation were successful, there would be
no difference between the results from the observational study and from the
target trial (had it been conducted).
In this chapter, we have explored three conditions—exchangeability, pos-
itivity, consistency—that help equate an observational study with a (con-
ditionally randomized) target experiment. When these conditions hold, we
can apply the methods described in the previous chapter—IP weighting or
standardization—to compute causal effects from the observational data.Fine Point 3.6 describes how to use
observational data to compute the
proportion of cases attributable to
treatment.
Therefore “what randomized experiment are you trying to emulate?” can
be a key question for causal inference from observational data. For each causal
effect that we wish to estimate using observational data, we may (i) specify
the target trial that we would like to, but cannot, conduct, and (ii) describe
how the observational data can be used to emulate that target trial. Specifying
the target trial, and therefore the causal effect of interest, requires specifying
key components of the trial’s protocol: eligibility criteria, interventions (or, in
general, treatment strategies), assignment, outcomes, start and end of follow-Hern´an and Robins (2016) speci-
fied the key components of the tar-
get trial. The acronym PICO (Pop-
ulation, Intervention, Comparator,
Outcome) is sometimes used to
summarize some of those compo-
nents (Richardson et al. 1995).
up, and causal contrasts.
Therefore, a valid emulation of the target trial requires that the observa-
tional dataset includes sufficient information to identify eligible individuals,
assign them to groups defined by the interventions they receive, and ascer-
tain their outcomes during the follow-up. For example, to estimate the causal
effect of heart transplant, we first specify the components of the protocol of
the target trial, and then try to emulate each of them using the observational
data. Such explicit emulation of a target trial improves causal inference from
observational data by making the interventions, and therefore the causal ques-
tion, well-defined (see Chapter 22 for an extended discussion of the target trial
framework). Once the causal question is well-defined via a target trial, investi-
gators can focus on whether and how conditional exchangeability across groups
can be achieved.When we are concerned that as-
suming conditional exchangeability
may not be reasonable given the
available data, we can consider
alternative identifying assumptions
(see Chapter 16) or perform sensi-
tivity analyses.
All of the above assumes that the interventions of interest are sufficiently
well-defined to translate them into a hypothetical experiment. But what can
we do when, based on current scientific knowledge, the causal question cannot
be translated into a target trial? As an example, consider “the causal effect
of weight loss” on mortality in individuals who are obese and do not smoke
at age 40. Because this causal question is somewhat vague, we replace it by
one that specifies the actual intervention that would be implemented to bring
about weight loss. For example, we could specify and emulate a target trial
about inducing weight loss via, say, exercise or diet.
In contrast to the conceptualization of causal inference from observational
data as a target trial emulation, some authors view “the causal effect of A on
Y ” as a well-defined quantity regardless of what A and Y stand for (as long as
A temporally precedes Y ). Often the argument goes like this:For some examples of this point of
view, see Pearl (2009), Schwartz
et al (2016), and Glymour and
Spiegelman (2016).
Requiring well-defined counterfactual outcomes via a target trial
imposes severe restrictions on the causal questions that can be
asked. Suppose we study weight loss and heart disease using ob-
servational data. We may not be able to characterize precisely the
causal effect of weight loss, but is that really so important if indeed
some causal effect exists? There is value in learning that many
deaths can be prevented if obese people, somehow, lost weight,